Agent frameworks & tooling

  • Munin lets research groups question papers on their own servers paper notes → — Its released software includes tests you can rerun after changing the model or paper collection.

    • An agent is a model running a multi-step task with tools. Munin searches papers, reads full passages and attaches citations to its answers.
    • The harness is the software around a model that runs its tools, files and commands. The Apache-2.0 repository combines Docker Compose, Qdrant document search and a Neo4j citation graph.
    • On 199 literature questions, Qwen3.8-27B answered 87.4% correctly with the full system, versus 38.7% unaided and 21.1% with five retrieved documents.
    • That full-system result permits web access. Blocking outbound requests lowers the same model’s accuracy to 66.3%, separating local-document usefulness from outside help.
    • You supply a model through an OpenAI-compatible endpoint. The reference setup uses two RTX 5090 cards with 32 GB each and a 68,500-paper collection. All groups share access to that paper collection.
    • Caveat: A fresh installation starts with an empty collection, so the published accuracy depends on rebuilding the relevant corpus and evaluation setup. (arXiv 2610.23480)

  • Superlite builds from a spec with fewer approval stops — The plugin is worth testing when you already know the requirements and want independent coding tasks to run together.

    • The MIT-licensed code records decisions, assigns files and interfaces to tasks, and uses hooks to keep its coordinator from editing project code.
    • The author reports one 16.1-minute build passing all 177 hidden tests. Earlier Superpowers runs averaged 33.6 minutes with its stock workflow and 23.5 minutes with added coordination. Each baseline average covers three runs.
    • That build used 21 agents, with up to 18 working simultaneously. Claude Code installation requires Git, Bash and Python 3.10 or newer.
    • Run it separately from Superpowers because their startup instructions and skill names conflict. The documented workflow still requires permission to push changes.
    • Caveat: The repository reports potentially higher cost and weaker extra robustness results, passing four of seven checks versus about 5.7 for Superpowers. (HN 99 points · 41 comments)

  • A WhatsApp reader lets assistants search your messages — It offers a small personal-message connector, with hosting and account risks worth understanding before linking a phone.

    • MCP is the standard for connecting agents to tools. This MIT-licensed server exposes chat listing, message retrieval and full-text search, while leaving message sending disabled.
    • Messages and login state live in SQLite inside your Cloudflare account. Connecting an assistant also makes retrieved messages available to that assistant’s model provider.
    • The current README requires Workers Paid and Node.js 20 or newer, despite the HN headline calling it free. Initial history import can exceed free-plan processing limits.
    • One deployment serves one owner. Imported history may be incomplete, and media appears as labels without downloaded files or transcription.
    • Caveat: The connector uses an unofficial WhatsApp Web library against WhatsApp’s terms, so read-only use still carries account risk. (HN 113 points · 65 comments)

  • LlamaIndex releases harder document tests with exact expected answers — ExtractBench Pro gives document-processing teams a downloadable way to test charts, drawings and rules spread across pages.

    • A benchmark is a fixed set of tasks for comparing models. The dataset contains 100 documents, 2,622 pages and 144,141 expected values across 86 document types.
    • Its six challenge groups cover document conventions, company vocabulary, counting rules, drawn relationships, charts and long chart sheets. Documents range from one to 83 pages.
    • The evaluator checks extracted values and their supporting locations. The headline document score requires every value to be correct, exposing failures that an average field score can conceal.
    • The launch tests allow 30 minutes per document. Failures and timeouts score zero for both extracted values and their supporting evidence.
    • Caveat: Every document is synthetic, with fictional values fixed before rendering, so success still needs checking against your actual customer files. (X @jerryjliu0)

  • TokenEconomizer learns where coding assistants waste money paper notes → — Its method targets repeated setup work and stalled attempts across successive tasks in the same repository.

    • The controller adds short instructions learned from completed work, hides repeatedly unused tools, and redirects stuck attempts before considering termination.
    • Across 277 coding tasks in six repositories, five assistant/model combinations and three random seeds, reported average cost fell 18.4–30.9%. Accounting includes controller requests and discounts for reused prompts.
    • First-attempt success improved in four combinations. Claude Code with Opus 4.8 fell from 74.0% to 73.0%, so the quality trade-off remains measurable.
    • Additional tests cut web-development costs by 16.0–21.2% and security-auditing costs by 6.4–19.6%. Periodic checks restore hidden tools to catch mistaken removals.
    • Caveat: Its learned stopping budget uses success feedback from the test grader, so production adoption needs an equally trustworthy signal of completed work. (arXiv 2610.22732)

  • Arcade measures when searching for tools cuts the bill paper notes → — The results help decide whether your assistant should load every tool description or find tools as needed.

    • Three discovery and execution tools expose an eligible catalog of about 8,200 tools. A separate authorization tool brings the client-visible interface to four tools and 993 tokens.
    • The gateway checks user consent and allowed tools before execution. On 7,961 ToolRet queries, every required tool appeared in the top ten 45.2% of the time, versus 28.8% for keyword search.
    • Against a fixed list of 100 tools, input fell 17.2% with Codex and 39.6% with Claude Code. Giving each client equal weight, bills fell 13.6% across 14 runs, with a 95% confidence interval of 7.4–18.4%.
    • Short lookups across catalogs of eight to 37 tools used 22.0% more input at the median. Each search adds another conversation turn.
    • Caveat: This is Arcade’s own platform study, and the Claude Code condition also shortened large tool results, so tool search alone cannot explain every saving. (arXiv 2610.23707)

  • SchemaShield-Lite repairs broken requests after tool descriptions change paper notes → — Validation followed by a constrained retry gives small-model deployments a concrete reliability method to test.

    • The software checks the selected tool and required arguments. Missing tool names trigger selection from visible tools before repair, without revealing the test’s correct tool.
    • On Qwen3.5-9B, accuracy rose from 71.7% to 81.3% across 300 dialogue cases, and from 58.0% to 80.5% across 200 function-request cases.
    • It recovered 19 of 22 and 33 of 45 failures caused by interface changes among previously correct cases. Those cases showed zero observed repair harms.
    • The public repository supplies a research scaffold. Tests outside the Qwen family show weaker gains when models fail to name a usable tool.
    • Caveat: The changed descriptions, renamed arguments and distracting tools were designed by the authors, so these results measure controlled interface changes rather than production reliability. (arXiv 2610.23342)

Models & research

  • Continued: Independent tests measure DeepSeek Flash’s speed and answer costs — day 2 of coverage (base specs in yesterday’s digest).

    • Artificial Analysis adds outside measurements to the release covered yesterday. The Max reasoning setting scores 45 on its Intelligence Index. Version 4.3.2 combines ten tests of coding, knowledge, document work and other tasks.
    • Through DeepSeek’s API, the page reports 227.1 output tokens per second and 1.09 seconds until the first token. Those are hosted-service measurements.
    • The index run generated 210 million output tokens, versus a 140-million median among comparable models. Its reported weighted average cost is $0.26 per index task.
    • These measurements give you speed, verbosity and task-cost baselines to compare against your own coding runs, beyond yesterday’s vendor-reported improvements.
    • Caveat: The index score and API speed describe this test setup, so neither establishes how well the model will perform on your workload or hardware. (HN 152 points · 54 comments)

  • HybridAgent mixes internal calculations with text to use tools paper notes → — The released models explore doing intermediate work without writing every step as words, while retaining ordinary Python and search requests.

    • The October 13 paper trains Qwen-based models at four, eight and 14 billion parameters. The repository includes a custom engine and downloadable eight- and 14-billion-parameter versions.
    • Its KV cache, a model’s stored working memory for a prompt, preserves internal vectors between tool interactions. A text transcript alone loses that state.
    • On the 14-billion-parameter comparison, GAIA task accuracy reaches 52.4%, versus 43.7% for the matched text-training method. WebWalker reaches 49.5%, versus 40.5%.
    • Estimated answer-generation work falls from 66.9 to 25.1 trillion arithmetic operations per question, a 2.7-fold reduction. That estimate excludes initial prompt processing, context-dependent attention and external tools.
    • Caveat: The method needs its custom generation software, and the estimated compute saving is not a measured reduction in an end-to-end service bill. (HN 407 points · 199 comments · arXiv 2610.18861)

Industry

  • Imprint gives narrowly eligible code changes a faster review path — Its published eligibility rules offer a concrete example of reducing review queues while preserving human review for sensitive work.
    • Median review time rose from 33 to 77 minutes over two years as average merged requests increased fourfold. The new policy currently covers about 10% of changes.
    • Eligibility requires fewer than 1,000 changed lines, no high- or medium-priority automated findings, and at least ten prior merged changes in every affected service.
    • The service-count thresholds are listed as one for frontend and three for backend. The post’s “fewer than” wording needs clarification at the frontend boundary. Shared libraries are excluded. Repositories can also exclude sensitive paths.
    • A build-controlled check determines eligibility, rather than a label people can edit. The review assistant can approve or ignore a request, limiting automated comment churn.
    • Caveat: The author says it is too early to report the policy’s effect, and existing compliance controls still require independent human review in some repositories. (lobste.rs)
All gathered items - what was cut and why (6)